Crash Risk Analysis of Non-Motorists using Interpretable Tabular Deep Learning

Abstract

The study investigates injury severity patterns among non-motorists (e.g., pedestrians and bicyclists) in Illinois using interpretable machine learning models applied to crash data from 2021 to 2023. The dataset included variables such as roadway design, environmental context, crash characteristics, and driver behavior. To address class imbalance in injury severity labels, synthetic oversampling was employed. Feature selection was performed using Random Forest and XGBoost, followed by training TabNet and FT-Transformer models. TabNet achieved the highest accuracy under the balanced training configuration and showed the strongest performance on the original imbalanced dataset across precision, recall, and F1 score, while also providing intrinsic interpretability through sequential attention-based feature selection. SHAP analysis revealed that temporal and environmental factors significantly influenced early model decisions, with behavioral and infrastructure-related variables impacting subsequent stages. The findings emphasize optimizing lighting, enhancing traffic control, and addressing time-specific crash risks. The study underscores the potential of interpretable tabular deep learning in traffic safety analytics, advocating for integration with contextual, geospatial, and behavioral data for targeted interventions.